Research 0662: Vulkan Motion Lavapipe Parity¶

Question¶

Can the always-on CPU↔Vulkan/lavapipe parity matrix cover the integer motion family again without weakening tolerances?

Findings¶

motion_v2_vulkan drifted from CPU on the Netflix normal pair with max_abs_diff = 2.624e-3 for both VMAF_integer_feature_motion_v2_sad_score and VMAF_integer_feature_motion2_v2_score.
CPU integer_motion_v2.c::mirror maps the high edge as 2 * size - idx - 2. The CUDA, SYCL, and Vulkan motion_v2 kernels instead used 2 * size - idx - 1, matching stale ADR-0193 prose rather than the current CPU reference.
Direct motion_vulkan still crashes on lavapipe in the Mesa llvmpipe driver, but the newer integer_motion_vulkan twin completes and matches CPU.
integer_motion_vulkan had debug=false by default, unlike CPU, CUDA, and legacy Vulkan motion, so parity could not compare the raw integer_motion metric until that default was restored.

Validation¶

All commands ran inside the vmaf-dev-mcp container unless noted otherwise.

VK_LOADER_DRIVERS_SELECT="*lvp*" \
LD_LIBRARY_PATH=/tmp/vmaf-build-vulkan-lvp/src \
.venv/bin/python scripts/ci/cross_backend_vif_diff.py \
  --vmaf-binary /tmp/vmaf-build-vulkan-lvp/tools/vmaf \
  --reference python/test/resource/yuv/src01_hrc00_576x324.yuv \
  --distorted python/test/resource/yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 --feature motion_v2 --backend vulkan --places 4

Result: both motion_v2 metrics had max_abs_diff = 0.0.

VK_LOADER_DRIVERS_SELECT="*lvp*" \
LD_LIBRARY_PATH=/tmp/vmaf-build-vulkan-lvp/src \
.venv/bin/python scripts/ci/cross_backend_vif_diff.py \
  --vmaf-binary /tmp/vmaf-build-vulkan-lvp/tools/vmaf \
  --reference python/test/resource/yuv/src01_hrc00_576x324.yuv \
  --distorted python/test/resource/yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 --feature motion --backend vulkan --places 4

Result: integer_motion, integer_motion2, and integer_motion3 all passed with max_abs_diff = 1e-6.

VK_LOADER_DRIVERS_SELECT="*lvp*" \
LD_LIBRARY_PATH=/tmp/vmaf-build-vulkan-lvp/src \
.venv/bin/python scripts/ci/cross_backend_parity_gate.py \
  --vmaf-binary /tmp/vmaf-build-vulkan-lvp/tools/vmaf \
  --reference python/test/resource/yuv/src01_hrc00_576x324.yuv \
  --distorted python/test/resource/yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 --backends cpu vulkan \
  --features adm ciede float_adm float_ansnr float_moment motion motion_v2 \
             float_motion float_ms_ssim float_ms_ssim_lcs float_psnr \
             float_ssim float_vif psnr psnr_hvs ssimulacra2 vif \
  --gpu-id "vulkan:0x10005:0x0" \
  --json-out /tmp/parity-gate-report/parity.json \
  --md-out /tmp/parity-gate-report/parity.md

Result: full lavapipe matrix passed with motion and motion_v2 included.

Limitation¶

An ad hoc CUDA+SYCL+Vulkan mixed build in /tmp generated the edited motion_v2_score.fatbin and integer_motion_v2_sycl.o, but failed at final link with pre-existing Vulkan private-symbol unresolved references in that LTO configuration. The focused Vulkan build and parity gate passed.